Detection of Diabetic Retinopathy using Convolutional Neural Network

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 09 Issue: 06 | June 2022 www.irjet.net p-ISSN: 2395-0072
© 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 1169
Detection of Diabetic Retinopathy using Convolutional Neural Network
J.Divya1, Roma B Das2, Shreya R Mummigatti3, Shweta Gudur4, Dr. Jagadeesh Pujari5
1234Student, Dept of Information Science Engineering, SDMCET, Dharwad, Karnataka, India.
5Professor, Dept of Information Science Engineering, SDMCET, Dharwad, Karnataka, India.
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract – Diabetic Retinopathy is a micro vascular
disorder that occurs due to the effect of diabetes that causes
vision damage to the retina, eventually leads to blindness.
Early detection of diabetic retinopathy protects the patients
from losing their vision. Detecting diabetic retinopathy
manually consumes lot of time and patients have to undergo
series of medical examinations. Hence, an automated system
helps in an easy and quick detection of diabetic retinopathy.
The proposed system inputs the image, extracts the features
such as micro aneurysms, exudates and haemorrhages and
classifies input image using convolution neural network as
normal or diseased based on its severity level. The result
obtained from the proposed methodgivesanaccuracy valueof
83%. And the system also predicts the severity level as mild,
moderate, proliferate and severe.
Key Words: Diabetic Retinopathy (DR), Convolutional
Neural Network(CNN).
1. INTRODUCTION
There are approximately ninety three million people
suffering with diabetic retinopathy worldwide.Thisnumber
seems to increase exponentially day by day. DR involves
different degrees of micro aneurysms, hemorrhages and
hard exudates in the peripheral retina. There exists various
effective treatments that would reduce the development of
the disease provided it is diagnosed intheinitial stagesitself.
The development of DR is a gradual process. There are
various symptoms of DR such as fluctuating and blurred
vision, dark spots and sudden vision loss. A web application
is implemented in this paper whichautomaticallydetectsDR
when an image is uploaded and is capable of classifying the
images based on the severity levels.
1.1 Literature Survey
Abnormality Detection in retinal images using Haar wavelet
and First order features [1]. A machine learning technique
has been a reliable one. The method classifies abnormality
detection in retinal images using Haar wavelet and First
order features. Diaretdb0 and Diaretdb1 are used as the
databases. A comparativestudyofDecisiontreeclassifierand
a KNN classifier is performed. Encouraging results were
found with the classification accuracy of 85%withK-Nearest
Neighbor classifier and 75% accuracy with decision tree
classifier. An automated detection of DR is done on real time
basis. The system proposed in this paper is computationally
inexpensive. The retinal image is classified as healthy or
diseased based on the application of haar wavelet and first
order histogram features.
Image processing technique are proposed here which
introduces a computer assisteddiagnosisbasedonthedigital
processing of retinal images.[2] in order to help people
detecting diabetic retinopathy in advance. The main goal is
to automatically classify the grade of non-proliferative
diabetic retinopathy at any retinal image. For that, an initial
image processing stage isolates blood vessels, micro
aneurysms and hard exudates in order to extract features
that can be used by a support vector machine to figure out
the retinopathy grade of each retinal image. Several image
pre-processing techniques have also beenproposedinorder
to detect diabetic retinopathy. However, despite all these
previous works, automateddetectionofdiabeticretinopathy
still remains a field for improvement. Thus, this paper
proposes a new computer assisted diagnosis based on the
digital processing of retinal images in order to help people
detecting diabetic retinopathy in advance.
An automated computer aided system is proposed in [3] for
the detection of DR using machinelearning hybridmodelsby
extracting the features like micro aneurysms, hemorrhages
and hard exudates. Theclassifierusedinthisproposedmodel
is the hybrid combination of SVM and KNN. Early medical
diagnosis of DR and its medical cure is essential to prevent
the severe side effects of DR. Manual detection of DR by
ophthalmologist takes lot of time and the patients need to
suffer lot. Hence, a system like this which is automated can
help in detection of DR quickly and we can easily follow up
treatment to avoid further effect to the eye.
A robust automatedsystem[3]isproposedwhichdetectsand
classifies the different stages of DR. The raw fundus images
are processed for removing noise and converting these
images into gray images using the steps for preprocessing to
ensure easier post processing. The optic disc and retinal
nerves are segmented. The features extracted are taken as
input parameters for the classification model for classifying
the images. The retinal nerves and optic disc are segmented,
and using Gray Level Co-occurrence Matrix (GLCM) method
the features are being extracted. An UI is implemented that
contains a push button to load an image which is used to
upload the input image then by using the radio button
present in the classifier is to be chosen by the user for the
classification process. After the classifier is chosen the
processing button is pressedwhich will processandshowall
the processed images with the result of the classification as
the stage of the Diabetic Retinopathy predicted.

An Interpretable Ensemble DeepLearningModelforDiabetic
Retinopathy Disease Classification [5] is proposed to detect
the presence of DR using multiple well – trained deep
learning models. The models makes use of various deep
learning model that are integrated with the help of Adaboost
algorithm. The weighted class activation maps (CAMs) are
used to demonstrate suspected position of lesions. There are
eight transformation ways introduced that helps in data
augmentation. The integrated model shows prominent
results when compared to the other individual deep leaning
models
1.2 Proposed System
In this paper a web application is introduced for the
detection of DR where the user has to first register using
valid credentials. The user then uploads the images on the
successful completion of registration. The process of
prediction is carried out by the model based on the input
image provided by the user. The model highlights the usage
of an eighteen layers deep ResNet-18, a CNN model. CNN
diagnoses from digital fundus images and accurately
classifies the severity of DR. CNN architecture and data
augmentation help to identify features involved in the
classification. The raw data is pre-processed to make it
suitable for machine learning model. The selected images
undergo segmentation on parameters such as presence of
blood vessels, Micro aneurysms,Hemorrhagesandexudates.
After the segmentation process the feature extraction
process takes place. The feature extractionstepresultsinthe
classification of normal and abnormal image. The normal
classifier indicates the absence of diabetic retinopathy. The
Abnormal classifier indicates the presence of diabetic
retinopathy and predicts the severityrateasmild,moderate,
proliferate or severe. The resultant prediction is shown in
the web page.
Fig 1: Architectural Flow
2. IMPLEMENTATION
2.1 Dataset
The primary data sourcefortheproposedsystemiscollected
from kaggle. The dataset contains 8596 images. The images
in the dataset are further classified into 5 categories out of
which, 3650 are normal 1392 are mild, 1972 are moderate,
708 are proliferate and 874 are severe respectively.
a) No_DR b)Mild c)Moderate
d)Proliferate e)Severe
Fig 2: Image Dataset
2.2 Data Augmentation
The data frames consists of training and testing sets of data.
Scikit-Learn is used to spilt the dataset out of which twenty
percent is spilt to testing and eighty percent is spilt to
training. Further two data generators are created each of
which is used for training and testing respectively. The
original images from the dataset are rotated and rescaled.
The training set is further divided into three namely train,
test and validation. Validation is used throughout the
training process. The process of rescaled of the images is
carried out in the testing set, where the images are rescaled
to 256 X 256 pixels and split into thirty two batches each.
2.3 Convolutional Neural Network
CNN takes the images from the training dataset consistingof
5 different classes namely normal, mild, moderate,
proliferate and sever. CNN consists of a dense fully
connected neural network having input, hidden and output
layer. All the neurons in CNN are fully connected. The input
image is fed in CNN where feature extraction is carried out.
The pooling layer compressesthefeaturemap. Thesefeature
maps are flattened and fed to the dense fullyconnectedlayer
before the classification. In general each of the CNN layer
learns filters in the increasing order of their complexity.The
first layer learns the basic feature extraction. The middle
layer learns about the detection of the objects and the final
layer learns to detect the object.

Fig 3. CNN Architecture
2.4 Residual Network
As CNNs grow deeper, vanishing gradient tend to occur
which negatively impact network performance. Vanishing
gradient problem occurs when the gradient is back-
propagated to earlier layers which results in a very small
gradient. Residual Neural Network includes "skip
connection" feature which enables training of 152 layers
without vanishing gradient issues. Resnet works by adding
"identity mappings" on top of the CNN.. ImageNet contains
11 million images and 11,000categories.ImageNetisusedto
train Res Net deep network.
Fig 4. ResNet-18
3. RESULTS AND CONCLUSIONS
3.1 Performance Requirements
Table I. Performance Index
The commonly used performancemeasurementsinmachine
learning are accuracy, precision and recall.
Accuracy = (TP + TN )/ TP +TN + FP + FN
Precision = TP/ (FP+ TP)
Recall = TP/ (FN + TP)
where TP is True positive, TN is True negative, FP is false
positive and FN is False negative
Fig 5. Predicted Output
3.2 Conclusion
Detecting diabetic retinopathy at an early stage is very
significant to prevent from causing severe blindness. The
symptoms that can be seen early for causing diabetic
retinopathy are blurred vision, darker areas of vision, eye
floaters and difficulty in recognizing colors. From the
extracted features of the image, the grade of the disease can
be classified as normal or abnormal. Also detecting and
diagnosing the disease in initial stage can help patientsfrom
blindness and can decrease the effects to the eye. In the
proposed system, the CNN model used is capable of
identifying features from the images provided. CNN is very
useful for clinicians of DR when datasets as well asnetworks
improve continuously to provide classification that can be
done in real-time.
REFERENCES
[1] S. Giraddi, S. Gadwal and J. Pujari,"Abnormalitydetection
in retinal images using Haar wavelet and First order
features," 2016 2ndInternational ConferenceonApplied and
Theoretical Computing and Communication Technology
(iCATccT), 2016.
[2] R, Nithya B , Reshma J , Ragendhu S and Sumithra M D
“Diabetic Retinopathy Detection using Machine Learning”
International Journal of Engineering Research &Technology
(IJERT), ISSN: 2278-0181 IJERT, Vol. 9 Issue 06, June-2020.

[3] S. Chaudhary and H. R. Ramya, "Detection of Diabetic
Retinopathy using Machine Learning Algorithm," 2020IEEE
International Conference for Innovation in Technology
(INOCON), 2020.
[4] K. Bhatia, S. Arora and R. Tomar, "Diagnosis of diabetic
retinopathy using machine learning classification
algorithm," 2016 2nd International Conference on Next
Generation Computing Technologies (NGCT), 2016.
[5] E. V. Carrera, A. González and R. Carrera, "Automated
detection of diabetic retinopathy using SVM," 2017 IEEE
XXIV International Conference on Electronics, Electrical
Engineering and Computing (INTERCON), 2017.
[6] Y. S. Kanungo, B. Srinivasan and S. Choudhary,"Detecting
diabetic retinopathy using deep learning," 2017 2nd IEEE
International Conference on Recent Trends in Electronics,
Information & Communication Technology (RTEICT), 2017.

Detection of Diabetic Retinopathy using Convolutional Neural Network

Recomendados

Recomendados

Mais conteúdo relacionado

Mais procurados

Mais procurados (20)

Semelhante a Detection of Diabetic Retinopathy using Convolutional Neural Network

Semelhante a Detection of Diabetic Retinopathy using Convolutional Neural Network (20)

Mais de IRJET Journal

Mais de IRJET Journal (20)

Último

Último (20)

Detection of Diabetic Retinopathy using Convolutional Neural Network